Structured Falsification (结构化证伪法)#

A five-step reasoning framework that forces rigorous disconfirmation before arriving at conclusions. Designed for AI agents and LLMs to produce concise, high-confidence outputs on complex tasks.

Core principle: Show conclusions, not derivation. The agent runs the full five-step process internally but outputs only the final ranked conclusions with confidence levels and key risks. Verbose reasoning is a sign the framework wasn't applied rigorously enough — tighten the analysis, don't expand the output.

When to Auto-Trigger#

• Multiple competing options with no clear winner
• User asks "should I…", "which one…", "what about…", "evaluate…", "compare…"
• Investment target analysis, due diligence, competitive assessment
• Technology selection, architecture decision, vendor evaluation
• Any task where the cost of a wrong answer is high

The Five Steps (internal process)#

Step 1: Decompose Value Nodes#

Map the problem space. Identify where real value / risk / leverage sits.

• What does this problem actually need? (Not surface requirements — underlying drivers)
• Map key entities and their relationships (supply chain / dependency graph / stakeholder map)
• Classify each node: critical vs. nice-to-have vs. irrelevant

Step 2: Falsify Each Candidate (core step)#

For every option / target / claim, run:

1. Surface logic: Why does the market / conventional wisdom support this?
2. Challenge: Where is the logic fragile? Causal chain breaks? Concept substitution? Hidden assumptions?
3. Verdict: Rate association strength — direct / indirect / tangential

Output: a falsification table (internal) with columns: Candidate | Surface Logic | Challenge | Verdict

Step 3: Identify True Beneficiaries / Best Options#

Apply priority filters:

1. Infrastructure / tooling (selling shovels during a gold rush) → highest certainty
2. Core technology owners (commercializable IP) → highest upside
3. Application layer (using tech to cut costs / add features) → value capture may be limited

Step 4: Stress Test Survivors#

For each surviving candidate:

• Is the causal chain A→B→C fully intact at every link?
• Is there actual evidence? (Business data, orders, customers, benchmarks)
• If this logic fails, how bad is the downside?
• Any show-stopper that eliminates this candidate entirely?

Step 5: Rank and Conclude#

• Sort by certainty (High / Medium / Low)
• Attach to each: one-line logic, key assumption, core risk
• Explicitly flag "not recommended" items with reasons
• One-sentence bottom line

Self-Correction Checklist#

Before producing output, verify:

• Did I falsify hard enough? If every candidate survived, the filter is too loose.
• Did I confuse "good company" with "good thesis"? Logic > quality.
• Am I hedging excessively? Pick a direction. Uncertainty should be flagged, not hidden behind "it depends".
• Did I anchor on the first plausible answer? Force a search for disconfirming evidence.
• Is my output concise? If the conclusion section exceeds 50% of the total output, re-tighten.

Output Format#

Only output the following. No step-by-step narration. No "let me think about this".

text

## 结论

| # | 候选 | 判断 | 确定性 | 核心逻辑（一句话） | 关键假设 | 主要风险 |
|---|------|------|--------|---------------------|----------|----------|
| 1 | ...  | ✅/⚠️/❌ | 高/中/低 | ... | ... | ... |
| 2 | ...  | ...  | ...    | ...                 | ...      | ...      |

**一句话总结：** ...

Domain Configurations#

Load domain-specific checklists when the context matches:

• Investment analysis → Read references/investment.md
• Technology selection → Read references/tech-decision.md
• Other domains → Use the generic framework above, or read a custom config from references/

To create a custom domain config, copy references/domain-template.md and fill in the sections.